a proposal for user-defined reductions in openmp · a proposal for user-deﬁned reductions in...

A proposal for User-Defined Reductions in OpenMP

A. Duran1, R. Ferrer1, M. Klemm2, B. de Supinski3, E. Ayguadé1

1BSC, 2Intel, 3LLNL

June 16th 2010

Outline

1 Motivation

2 UDR Design rationale

3 Declaring UDRs

4 Array reductions

5 C++ specific extensions

6 Conclusions

Duran et al. () UDRs in OpenMP June 16th 2010 2 / 28

Motivation

Outline

1 Motivation


3 Declaring UDRs

4 Array reductions


6 Conclusions


Motivation

Reductions in OpenMP 3.0

Current OpenMP supports reduction:basic scalar typessimple arithmetic operators (+,-,*,&,...)array reductions (Fortran only)min and max operators (Fortran only)

Users with other reductions must find their way manually:Using critical

Using atomic (when possible)Adding complex code to implement the reduction


Motivation

Reductions by hand

1 complex_t complex_mul ( complex_t a , complex_t b ) ;2

3 void example ( complex_t ∗array , s i z e _ t N) {4 complex_t prd = { 1 . 0 , 0 . 0 } ;5

6

7

8

9 #pragma omp parallel reduction ( prd )10 {11

12 #pragma omp for13 for ( s i z e _ t i = 0 ; i < N; i ++)14 prd = complex_mul ( prd , a r ray [ i ] ) ;15

16 }17

18

19 }


Motivation

Reductions by hand



6 i n t nthreads = omp_get_max_threads ( ) ;7 complex_t par t_prd [ nthreads ] ;8

9 #pragma omp parallel shared ( par t_prd ) private ( prd )10 {11 prd = { 1 . 0 , 0 . 0 } ;12 #pragma omp for13 for ( s i z e _ t i = 0 ; i < N; i ++)14 prd = complex_mul ( prd , a r ray [ i ] ) ;15 par t_prd [omp_get_thread_num ( ) ] = prd ;16 }17

18 for ( i n t t h r = 0 ; t h r < nthreads ; t h r ++)19 prd = complex_mul ( prd , par t_prd [ t h r ] ) ;20

21 }


Motivation

Not goodDrawbacks

More complex user codeError proneDoesn’t benefit from implementation improvements

SolutionAdd user-defined reductions to OpenMP


Motivation

User-defined reductions

Allow the user to inform OpenMP about new reductions by providing:A typeAn operation over that typeThe identity value for that operation and type

Presented to the OpenMP language committee (still on discussion)


UDR Design rationale

Outline

1 Motivation


3 Declaring UDRs

4 Array reductions


6 Conclusions


UDR Design rationale

Driving design goals

follow OpenMP directive-based philosophysupport all OpenMP base languages

but mantain a common syntax as much as possible

follow a declaration/usage patternAllow code re-useAllow efficient implementation

require associativity and commutativity


Declaring UDRs

Outline

1 Motivation


3 Declaring UDRs

4 Array reductions


6 Conclusions


Declaring UDRs

Declaring an UDRSyntax

C:1 #pragma omp declare reduction ( operator− l i s t : type ) [ c lause ]

C++:1 #pragma omp declare reduction ( [ template <template−params >] operator−l i s t : type−l i s t ) [ c lause ]

Fortran:1 !$omp declare reduction ( operator− l i s t : typename ) [ c lause ]

where clause is:1 identity ( expression | brace−i n i t i a l i z e r C/C++ | constructor [ ( argument−l i s t ) ] C++ )


Declaring UDRs

Example


3 #pragma omp declare reduction ( complex_mul : complex_t ) \4 identity ( { 1 . 0 , 0 . 0 } )5


9 #pragma omp parallel for reduction ( complex_mul : prd )10 for ( s i z e _ t i = 0 ; i < N; i ++)11 prd = complex_mul ( prd , a r ray [ i ] ) ;12

13 }


Declaring UDRs

Operator requisites

binary operators with compatible arguments with UDR typeReturn value either by return or parameter (*,&)

Return value priority: function return, left parameter, right parameter

Allow const and references(C++) Unary member functions

specified by prepending a dot to the operator name

(C++) valid overloaded operatorsassociativecommutativeavailable using base language “symbol“ lookup

both at declaration and usage


Declaring UDRs

Identity value

By default:C/Fortran Zero initialization

C++ C++ rules for value-initializationCan be overrided by the identity clause

The identity expression must evaluate always to the same value

An implementation can evaluate it one or more times


Array reductions

Outline

1 Motivation


3 Declaring UDRs

4 Array reductions


6 Conclusions


Array reductions

Array UDRs

DeclarationTypes can be prepended with [] that indicate that is going to be anarray UDR

One [] per dimensionDimension size is not fix at declaration

Operators can have additional integer parameters to get the actualsize

UsageArray UDRs can be applied to variables of:

Array typesPointer to array types

allows to ”recover” arrays through function calls

Support for VLA (pointers to) arrays is limited to C.


Array reductions

Example

1 const i n t N = 10;2 void vector_add ( i n t ∗A, i n t ∗B, i n t n ) ;3

4 #pragma omp declare reduction ( vector_add : i n t [ ] )5

6 void foo ( i n t ∗ a , i n t ∗ b , i n t n )7 {8 i n t v1 [N ] ;9 i n t (∗ v2 ) [N] = ( i n t (∗ ) [N ] ) a ;

10 i n t (∗ v3 ) [ n ] = ( i n t (∗ ) [ n ] ) b ; / / VLA ; Only v a l i d i n C11

12 #pragma omp for reduction ( vector_add : v1 , v2 , v3 )13 for ( . . . ) { . . . }14 }


C++ specific extensions

Outline

1 Motivation


3 Declaring UDRs

4 Array reductions


6 Conclusions



Constructed indentities

The special constructor keyword can be used in the identityclause

Private copies will be initialized with a constructor instead ofby-assignment

1 #pragma omp declare reduction ( ∗ : Complex ) identity (constructor ( 1 . 0 , 0 . 0 ) )



Inheritance support

IdeaSupport the C++ philosophy of allowing to use methods defined overbase classes with derived classes

UDR of base classes can be used for derived classesOnly if it does not create object slicing

operator parameters must be pointers or references



Template UDRs

Allows to declare an UDR on all (or partial) instantiations of templatetype

1 #pragma omp declare \2 reduction ( template <class T> std : : l i s t <T > : : merge : s td : : l i s t <T> )3

4 void foo ( )5 {6 s td : : l i s t < int > l i ;7 s td : : l i s t < f loa t > l f ;8

9 #pragma omp parallel for reduction ( s td : : l i s t < int > : : merge : l i )10 reduction ( s td : : l i s t < f loa t > : : merge : l f )11 for ( . . . ) { . . . }12 }



Dot syntax

Simplify qualification for C++ by doing auto-qualification based on theUDR type

1 /∗ #pragma omp dec lare \2 reduc t ion ( template <c lass T> . merge : s td : : l i s t <T> ) ∗ /3

4 #pragma omp declare \5 reduction ( . merge : s td : : l i s t < int > , s td : : l i s t < f loa t > )6

7 void foo ( )8 {9 s td : : l i s t < int > l i ;

10 s td : : l i s t < f loa t > l f ;11

12 #pragma omp parallel for reduction ( . merge : l i , l f )13 for ( . . . ) { . . . }14 }


Conclusions

Outline

1 Motivation


3 Declaring UDRs

4 Array reductions


6 Conclusions


Conclusions

Conclusions

Proposal allows users to extend OpenMP with their own reductionoperations

aggregated typesarraystemplate types

Still in discussion for OpenMP 3.1We’re looking for user codes


Conclusions

The End

Thanks for your attention!


Conclusions

Min/Max

1 #pragma omp declare \2 reduction ( template <typename T> std : : min : T ) \3 identity ( s td : : numer ic_ l im i ts <T > : : max ( ) )4 #pragma omp declare \5 reduction ( template <typename T> std : : max : T ) \6 identity ( s td : : numer ic_ l im i ts <_T > : : min ( ) )


Conclusions

Some example

1 i n l i n e void poin tMin ( Po in t ∗a , Po in t ∗b )2 {3 i f ( a−>x < b−>x ) a−>x = b−>x ;4 i f ( a−>y < b−>y ) a−>y = b−>y ;5 }67 i n l i n e void pointMax ( Po in t ∗a , Po in t ∗b )8 {9 i f ( a−>x > b−>x ) a−>x = b−>x ;

10 i f ( a−>y > b−>y ) a−>y = b−>y ;11 }1213 #pragma omp declare reduction ( pointMax : Po in t ∗)14 #pragma omp declare reduction ( po in tMin : Po in t ∗)1516 void f indMinMax ( Po in tVec tor po in ts , index_ t n , Po in t∗ minPoint ,17 Po in t∗ maxPoint )18 {1920 #pragma omp parallel for schedule ( s t a t i c ) \21 reduction ( po in tMin : minPoint ) reduction ( pointMax : maxPoint )22 for ( index_ t i = 1 ; i < n ; i ++) {23 i f ( po in t s [ i ] . x < min . x ) minPoint−>x = po in t s [ i ] . x ;24 i f ( po in t s [ i ] . y < min . y ) minPoint−>y = po in t s [ i ] . y ;25 i f ( po in t s [ i ] . x > max . x ) maxPoint−>x = po in t s [ i ] . x ;26 i f ( po in t s [ i ] . y > max . y ) maxPoint−>y = po in t s [ i ] . y ;27 }28 }29 }


Conclusions

Gromacs1 void array_rvec_add ( rvec ∗a , rvec ∗b , i n t n )2 {3 for ( i n t i = 0 ; i < n ; i ++) {4 a [ i ] . xx += b [ i ] . xx ;5 a [ i ] . yy += b [ i ] . yy ;6 a [ i ] . zz += b [ i ] . zz ;7 }8 }9

10 void rea l_array_add ( r e a l ∗a , r e a l ∗b , i n t n )11 {12 for ( i n t i = 0 ; i < n ; i ++) a [ i ] += b [ i ] ;13 }1415 #pragma omp declare reduction ( array_rvec_add : rvec [ ] ) identity ( { 0 . 0 , 0 . 0 , 0 . 0 } )16 #pragma omp declare reduction ( ar ray_real_add : r e a l [ ] ) identity ( 0 . 0 )1718 . . .19 neg2 = mdatoms−>nenergrp∗mdatoms−>nenergrp ;2021 rvec (∗ pf ) [ mdatoms−>nr ] = ( rvec (∗ ) [ mdatoms−>nr ] ) f ;22 rvec (∗ p f s h i f t ) [ SHIFTS ] = ( rvec (∗ ) [ SHIFTS ] ) p f s h i f t ;23 r e a l (∗pegcoul ) [ neg2 ] = ( r e a l (∗ ) [ neg2 ] ) egcoul ;24 r e a l (∗pegnb ) [ neg2 ] = ( r e a l (∗ ) [ neg2 ] ) egnb ;2526 #pragma omp parallel for reduction ( array_rvec_add : pf , p f s h i f t ) \27 reduction ( ar ray_real_add : pegcoul , pegnb )28 for ( b=0; b<nb ; b++)29 f u n c t i o n ( p f s h i f t , pf , pegcoul , pegnb ) ;3031 . . .


a proposal for user-defined reductions in openmp · a proposal for user-deﬁned reductions in...

Documents